Optimization of Virtual Loudspeakers for Spatial Room Acoustics Reproduction with Headphones
Total Page：16
File Type：pdf, Size：1020Kb
Article Optimization of Virtual Loudspeakers for Spatial Room Acoustics Reproduction with Headphones Otto Puomio * ID , Jukka Pätynen and Tapio Lokki ID Department of Computer Science, Aalto University School of Science, P.O. Box 13300, 00076 Aalto, Finland; jukka.patynen@aalto.fi (J.P.); tapio.lokki@aalto.fi (T.L.) * Correspondence: otto.puomio@aalto.fi; Tel.: +358-45-678-5213 Received: 31 October 2017; Accepted: 5 December 2017; Published: 9 December 2017 Abstract: The use of headphones in reproducing spatial sound is becoming more and more popular. For instance, virtual reality applications often use head-tracking to keep the binaurally reproduced auditory environment stable and to improve externalization. Here, we study one spatial sound reproduction method over headphones, in particular the positioning of the virtual loudspeakers. The paper presents an algorithm that optimizes the positioning of virtual reproduction loudspeakers to reduce the computational cost in head-tracked real-time rendering. The listening test results suggest that listeners could discriminate the optimized loudspeaker arrays for renderings that reproduced a relatively simple acoustic conditions, but optimized array was not significantly different from equally spaced array for a reproduction of a more complex case. Moreover, the optimization seems to change the perceived openness and timbre, according to the verbal feedback of the test subjects. Keywords: Spatial audio; Spatial sound reproduction; SDM; Headphone reproduction; Optimization 1. Introduction Spatial audio aims to reproduce a believable illusion for a listener being in a real acoustic space by electronic means [1]. Dozens of different ways exist to record or artificially create the spatial sound signals, which are further reproduced with an array of loudspeakers or with headphones [2]. For good spatial resolution, a high number of reproduction loudspeakers are often used in research facilities or in special venues, but such arrays are impractical in domestic or other daily listening environments. Therefore, the headphone reproduction of spatial sound is gaining interest. Commonly, the headphone-based spatial sound is implemented by virtualizing the reproduction loudspeaker array with the Head-Related Transfer Functions (HRTFs) [3]. In essence, HRTFs are applied to virtually position the sound sources around the listener. The resulting binaural rendering can sound very convincing at the best, but, for some users, the sound is localized inside the head. To achieve better externalization, head-tracking devices are often used to compensate the movement of the listener’s head and to keep the reproduced auditory environment stable [4,5]. This work studies the virtual loudspeaker positioning in headphone-based spatial sound reproduction. The sound rendering is based on the Spatial Decomposition Method (SDM) [6], which in essence analyzes directional information in spatial room impulse responses (RIRs). In brief, the SDM uses a compact array of microphones in a RIR measurement. Based on the time difference of arrivals between microphone pairs in short time windows, it estimates the direction of arrival (DOA) for each audio sample in the captured RIR. Therefore, the SDM allows a wide range of perceptual room acoustics studies, and it has been recently applied to study concert halls [7,8], studio control rooms [9], car cabins [10,11], as well as stage acoustics for musicians [12]. To reproduce the spatial sound analyzed with the SDM, audio samples in RIR are assigned to the loudspeakers of a given reproduction array. The assignment yields a number of sparse impulse Appl. Sci. 2017, 7, 1282; doi:10.3390/app7121282 www.mdpi.com/journal/applsci Appl. Sci. 2017, 7, 1282 2 of 16 responses equal to the number of loudspeakers used. The spatial sound is finally rendered by convolving sound signals with each of these sparse impulse responses. That is, the measured RIR is distributed spatially as convolution reverberators for a defined reproduction loudspeaker array. Although the direction of arriving sound is accurately estimated in the analysis, the practical limits in the real or virtual loudspeaker array introduce varying amounts of angular error to the spatial sound reproduction. In the worst case, the mismatch between original and synthesized DOA may change the spatial image of the acoustic space drastically. In theory, all angular errors could be avoided and the analyzed sound field could be reproduced perfectly by assigning each sample to its own loudspeaker. However, this kind of approach is physically infeasible to implement, especially for real room-acoustic conditions. One popular method to reduce the angular error in spatial sound synthesis is the Vector Base Amplitude Panning (VBAP) [13]. VBAP weights each sound sample between the three closest reproduction loudspeakers so that the resulting sound appears to arrive from the original direction. When applied to all samples in the RIR, the spatial image of the space is reproduced correctly. However, it is known that amplitude panning of samples corresponds to time-averaging of multiple HRTFs, leading to low-pass filter effect on the perceived sound [14]. Earlier studies employing SDM [7–12] have utilized a more straightforward synthesis method, namely Nearest Loudspeaker Synthesis (NLS), partially to circumvent the potential spectral issues with VBAP. NLS distributes each audio sample of a single monaural RIR to the nearest reproduction loudspeaker based on the estimated DOA information. Since only one reproduction loudspeaker at a time is involved in reproducing the RIR sample, this approach is free from the effects of HRTF averaging, but at the cost of increased angular error. The aforementioned studies have used predetermined physical loudspeaker arrays for reproduction. Similar studies could still benefit from increased fidelity of spatial sound reproduction either by increasing the number of loudspeakers or by using the existing loudspeakers more efficiently. However, in many cases, adding loudspeakers is less feasible than optimizing the existing physical setup. When the reproduction of spatial audio is substituted with headphone listening, as in this study, the optimization becomes even more sensible. In theory, headphones enable the use of practically an unlimited number of virtual loudspeakers in any given direction. However, since the computational cost is directly proportional to the number of spatial channels rendered for headphones, using a smaller number of virtual loudspeakers is preferred. This requirement justifies smarter allocation of resources, which, in this case, means optimized virtual loudspeaker positions. This paper presents a method of determining a room-specific virtual loudspeaker array that reproduces spatial sound perceptually more efficiently than a predetermined conventional array. The proposed method aims to minimize the directional error of NLS as well as to enhance labor-to-quality ratio of the rendering. In other words, spatial sound can be reproduced either more accurately with the same number of virtual loudspeakers, or at the same quality with a reduced channel number. These so-called optimized loudspeaker setups are compared with uniformly distributed setups to measure how recognizable are the differences in reproduction. The paper is structured as follows. First, Section2 outlines the structure of the position optimization system. Next, Section3 describes the listening test, followed by the results in Section4. Finally, more detailed analysis is presented with discussion in Section5 and wrapped up in Section6. 2. Virtual Loudspeaker Position Optimization The sound reproduction system used in this paper is similar to one presented by Tervo et al. [10]. The system performs SDM analysis and NLS reproduction as described in Section1, resulting in a sparse impulse response (IR) for each reproduction loudspeaker. Finally, those sparse IRs are convolved with the audio signals to create the spatial sound. The used version of the SDM also post-equalizes the loudspeaker IRs as the rapid channel changes of the IR cause whitening of the signal. It should be noted that the optimization of the reproduction loudspeaker positions is done Appl. Sci. 2017, 7, 1282 3 of 16 for SDM data before the convolution step. In other words, the optimization requires a spatial sound reproduction technique that has information on the spatial IR for each sound source, and thus cannot be applied to an arbitrary spatial sound reproduction technique. The NLS requires information on the reproduction loudspeaker positions in order to be able to distribute the samples properly. As mentioned in Section1, the positions are usually static and not changed according to the acoustics of the space being reproduced. The method presented in this section replaces these static loudspeaker positions with optimal ones that are calculated from the RIR of the room being reproduced. The system used in reproduction is therefore the same as described above except for the way the loudspeaker setup is determined. Figure1 outlines the position optimization process which is done in two steps in its most basic form. First, SDM samples including RIR pressure and directional metadata (azimuth and elevation) are weighted according to their energy as well as their spatiotemporal properties. Then, loudspeaker positions are initialized based on the calculated weights and the final positions are obtained by clustering weighted DOAs iteratively until convergence. The clustering is